On its own, no. The statistics most instructors see — P-values, point-biserials, Cronbach's alpha — describe how each item behaved for the whole class. They can't tell you anything about two particular students. But most item-analysis exports also contain a per-student response sheet, and that sheet can be screened for patterns that are hard to produce honestly. The important word is screened. What comes out is a list of things to look at, not a finding.
The signal that matters: shared wrong answers
Two strong students will agree on most correct answers. That's knowing the material. What they shouldn't do is repeatedly pick the same wrong option.
On a four-option item that two students both miss, the chance they chose the same distractor is roughly one in three — higher if one distractor is popular, lower if the class spread evenly. Across every item a pair both missed, you can add those chances up and get an expected number of matching wrong answers. When a pair's actual count is far above that — 9 matches where 3 were expected — the probability of that happening by chance is small enough to write down.
A good screen writes it down. It also corrects for the fact that a class of 54 has 1,431 pairs: with that many comparisons, one or two will look odd by luck alone, so the threshold has to be stricter than it would be for a single pair.
Secondary signals
- Timing. Exams finished in a third of the class's median time, or the hardest items answered faster than the easiest, when the export includes per-item timing.
- Answer-change bursts. A quarter of the exam changed in the last few minutes.
- Compromised items. A question that was answered correctly by 48% of students last semester and 94% this semester — with its discrimination collapsing at the same time — may be circulating.
- Identical response strings. Three or more students with the same answer on every item, wrong ones included.
What a flag is and isn't
A study group that memorized the same wrong fact looks, statistically, a lot like two students copying. So does a pair who shared a review sheet with an error on it. The numbers can't distinguish those stories; only you can, with seating charts, proctoring records, the students' prior work and a conversation.
That's why an integrity screen should never produce a "likelihood of cheating" score. It should show the evidence — items both missed, identical wrong choices, expected by chance, probability — in language you could read aloud to a student or an appeals committee, and then stop.
Protecting students while you look
If you send response data to any outside service, insist on three things: identifiers hashed before they leave your computer, with the lookup key staying with you; response-level data deleted on a short schedule; and findings delivered to you, not your institution. The data you're screening belongs to your students, and the person who decides what it means should be their instructor.
MapMyExam's Exam Integrity Screen is an opt-in add-on built on exactly those rules.